Anomaly Detection in Smart Grid Load Profiles via Variational Autoencoders

This paper proposes a smart grid load anomaly detection method based on Variational Autoencoder (VAE). By leveraging deep learning to analyze smart grid load data, the proposed method automatically learns the distribution characteristics of normal load profiles and identifies anomalies through reconstruction error analysis. Experimental results on the Smart Grid Stability dataset demonstrate that the VAE model outperforms traditional approaches such as Autoencoder and Isolation Forest, achieving a precision of 0.87, recall of 0.84, and Area Under the Receiver Operating Characteristic Curve (AUC) of 0.91. These results represent improvements of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$10-20 \%$</tex> over baseline methods. Furthermore, SHapley Additive exPlanations (SHAP) are used to interpret the model outputs and highlight the most influential features in anomaly detection, enhancing the explainability and practical applicability of the method in realworld smart grid monitoring scenarios.